Object Tracking And Detection Based On Active Learning

In the current research and applications of machine learning,models need a lot ofanually labeled data to train.A large number of annotation tasks in computer vision tasks not only burden researchers,but also the accuracy of labeling is not always satisfactory.Because it very difficult to obtain a large amount of manually labeled data,which allows great development space for active learning algorithms that require only a small number of labeled samples.The key to active learning lies in the strategy of selection.However,many strategies can only apply to specific areas,which has certain limitations and cannot be extended to other areas.There are no outstanding active leaning algorithms for object tracking and detection based on deep learning.In order to solve the problem,we designed different active learning strategies for two different tasks.For the object tracking task,the current popular siamese object tracking algorithms are training on the sample pairs.Therefore,we design an active learning strategy based on spatio-temporal consistency called STAL(spatio-temporal active learning strategy),which can make good use of time information.It can mine the samples with high information base on STAL strategy.We also adds self-paced learning to control the difficulty of training data in different periods.We will use the output of the pre-trained model to evaluate the spatial consistency of proposals and template.Experiments on several public datasets(for example,OTB-100,VOT-2015,VOT-2016 and UAV123)showed that STAL can learn better with less sample pairs and get a more accurate model.For the object detection task,not only the size of training data can affect the performance of the model,but also the distribution of training data.Since the training data of the object detection task is atomic,we can get the distribution of the data.So we propose an active learning strategy based on data distribution,which integrates with popular object detection algorithms.The strategy measures the distribution between the feature set of the labeled data and unlabeled data in the Hilbert space to select the most representative samples,and evaluates the uncertainty of samples based on entropy simultaneously.In our implements,we using the initialized model to extract the high-level features of samples each batch,and then selecting samples for training based on representative and entropy.Compare to other active learning strategies,we have achieved outstanding performance in the public PASCAL VOC and COCO datasets.